Magellan Final Report - Office of Science - U.S. Department of Energy

Magellan Final Report
Productivity. One advantage of cloud computing can be increased productivity. For users who need the
added flexibility offered by the cloud computing model, additional costs may be more than offset by the increased
flexibility. Furthermore, in some cases the potential for more immediate access to compute resources
could directly translate into cost savings. Section 12.5 discusses examples where this could occur. Potential
increases in productivity should be weighed against any up-front investments required to re-engineer software
or re-train staff.
Acquisition Costs. It is generally assumed that cloud providers pay significantly lower costs for hardware
than typical IT customers. This assumption is reasonable given the purchasing power a cloud provider can
exercise. Since cloud providers do not publish their acquisition costs, it is difficult to make direct comparisons.
However, DOE centers use competitive procurements or close partnerships to procure their systems. These
procurements are typically very large and command substantial discounts from both the integrator and
component providers. These discounts may not be quite as high as for warehouse scale datacenters, but
likely approach them. Furthermore, the purchasing contracts typically include acceptance criteria tied to
application performance which significantly reduces risks.
12.4 Historical Trends in Pricing
One argument often made in favor of commercial clouds is that it enables customers to automatically
reap the cost benefits of technology improvements. HPC customers have come to count on the “Moore’s
Law” improvements in computing capability which typically result in a doubling in capability per dollar
approximately every two years. As a result, DOE centers have historically delivered average improvements
in computing capability of 40%-80% per year with relatively flat budgets. However, this level of decreasing
costs for compute performance is not currently observed in the commercial cloud space. For example, the cost
of a typical compute offering in Amazon (m1.small) has fallen only 18% in the last five years with little or no
change in capability. This translates into a compound annual improvement of roughly 4%. Meanwhile, the
number of cores available in a typical server has increased by a factor of 6x to 12x over the same period with
only modest increases in costs. While there are examples of cost reductions in commercial cloud services,
such as free in-bound data transfers or lower storage cost for higher tiers, the overall pricing trends, especially
for computation, still show only modest decreases. The new systems coming online at the ALCF provide an
example of how DOE centers delivere increased computing capability with relatively flat budgets. The three
new IBM Blue Gene/Q systems at the ALCF will provide an estimated total HPL Peak of 8.4PF (based
on prototype Blue Gene/Q systems currently listed on the November 2011 Top500 list) at an anticipated
annual budget of $62M/year. This results in a Cost per TF to operate for 1 year of $7.4K, which is 13-14x
better than Amazon’s calculated cost (Section 12.2.4), and is an 11x improvement over the current ALCF
resource, Intrepid, which is around four years old.
12.5 Cases where Private and Commercial Clouds may be Cost
Effective
There are some cases where moving a workload to a private or public cloud offering can be cost effective.
We will briefly describe a few of these cases.
Unknown Demand. In the case of a new project or a new application area where the potential demand is
still poorly understood, it can be cost effective to use a commercial cloud while the demand is quantified and
understood. Once the level has been established, in-house resources can deployed with greater confidence.
Similarly, if a new project’s prospects are highly uncertain, clouds can be a cost effective option while the
long-term fate is determined. In both cases, savings are achieved by avoiding investments in unneeded ca-
125